Antibody-antigen interfaces can be predicted using unrelated monomeric protein data

Prediction of antibody-antigen interfaces is crucial for rational antibody design. Thus far the field is facing a data problem wherein we only have ca. 1,000 non-redundant antibody-antigen complexes in a highly complex problem. By contrast they are close to ca. 14,000 non-redundant single protein complexes. It is plausible to assume that the phenomena governing protein folding might follow similar principles as transient complex formation. Therefore, molecular information from the 14,000 monomers should inform the antibody-antigen complex formation. To test this hypothesis we train a simple neural network using the protein data and apply it to the antibody-antigen interfaces. We demonstrate that the network that was trained on protein data and had no notion of antibody sequence, performs better than the network trained on antibody data. Therefore internal atomic conformations can be used to extend the data for predicting protein-protein and thus antibody-antigen interactions.